Medical time-series data play a vital role in disease diagnosis but suffer from limited labeled samples and single-center bias, which hinder model generalization and lead to overfitting. To address these challenges, we propose DAAC (Discrepancy-Aware Adaptive Contrastive learning), a learnable multi-view contrastive framework that integrates external normal samples and enhances feature learning through adaptive contrastive strategies. DAAC consists of two key modules: (1) a Discrepancy Estimator, built upon a GAN-enhanced encoder-decoder architecture, captures the distribution of normal data and computes reconstruction errors as indicators of abnormality. These discrepancy features augment the target dataset to mitigate overfitting. (2) an Adaptive Contrastive Learner uses multi-head attention to extract discriminative representations by contrasting embeddings across multiple views and data granularities (subject, trial, epoch, and temporal levels), eliminating the need for handcrafted positive-negative sample pairs. Extensive experiments on three clinical datasets—covering Alzheimer’s disease, Parkinson’s disease, and myocardial infarction—demonstrate that DAAC significantly outperforms existing methods, even when only 10\% of labeled data is available, showing strong generalization and diagnostic performance. Our code is available at https://github.com/CUHKSZ-MED-BioE/DAAC.

Prediction of daily mental wellbeing holds profound implications for individual healthcare and societal stability. Previous studies have shown the potential of using individual's multimodal behavioral data collected through mobile devices to predict his/her daily mental wellbeing metrics, such as stress, mood, and anxiety. However, effectively capturing long-range dependencies in behavioral time series data while accurately representing the statistical distribution patterns of various behaviors over a certain period is a significant challenge. In this paper, we propose a daily mental wellbeing prediction model based on a Topic-Guided Self-Attention Network (TGSAN). This model utilizes self-attention mechanism to capture long-range dependencies from the behavioral data collected by mobile devices. We utilize a multi-granularity time encoding method to inject time information of different granularities (i.e., day and hour, or week and day) into the behavioral data, thereby enhancing the sensibility of the self-attention network to capture every individual's habitual cyclicality rhythm. Then, we introduce a neural topic model to analyze the statistical distribution characteristics of various behaviors in the monitoring period as behavioral distribution patterns for different individuals, and further propose a topic attention network to enhance the model's classification performance by guiding the weights of long-range dependencies features from the self-attention network with the derived topic information. Compared to state-of-the-art methods, the proposed TGSAN achieved superior performance on datasets that measure different mental health indicators (stress, mood, and anxiety), with F1 scores outperforming by 4.5% and 2.3% on the Crosscheck and StudentLife datasets, respectively, and accuracy outperforming by 3.3% on the GLOBEM dataset. Our study demonstrates the effectiveness and interpretability of combining self-attention mechanisms with neural topic model, for a better understanding of the relationship between different individuals’ behaviors and their mental wellbeing.

Rail corrugation is a type of track geometry irregularity that induces sharp noise and reduces the service life of track components. The traditional manual inspection method of corrugation detection requires high labor costs. In general, the rail corrugation status of subway lines can be divided into three types according to wavelength: short wave, medium wave and healthy. Corrugation with different degrees of deterioration has different values of wave depth. To achieve more efficient and intelligent inspection, this study proposes a multitask learning method to detect metro rail corrugation by using in-vehicle responses and noise data. To this end, we develop a portable vehicle-mounted device to collect car body acceleration and vehicle interior noise. Second, we propose a data fusion method based on a 1/3 octave general vibration level of acceleration and noise data. Third, we develop a multitask learning model based on convolution and attention techniques that have two task towers: corrugation wavelength classification and wave depth regression. Finally, we apply the real-world data to the proposed framework. The results show that compared with single-task learning, the proposed multitask learning has a better performance in corrugation detection tasks. The$F_{1}$score of wavelength classification increased from 0.762 to 0.971, an increase of 27%; the Mean Absolute Percentage Error of wave depth regression decreased from 12% to 4.96%, a decrease of 59%. In addition, the training time consumption can be reduced by up to 66%.